semantic layer bot

Setup

poetry install --with dev

To interact with real semantic layer

• Replace semantic_manifest.yaml with real manifest
• Remove the mocked dbt calls in agent.py file (_get_dbt_query_result() and _create_dbt_query())

Run indexing

poetry run python ingest.py

Use agent

from agent import graph

async for event in graph.astream_events({"messages": [("human", "how much was ordered in total in july")]}, version="v2", include_types=["chat_model", "tool"]):
    print(event['event'])
    print(event['data'])
    print("\n\n" + "-" * 80 + "\n\n")

TODO

• Is retrieval even needed? If so:
  • Use a real vecstore (not a local one)
  • Use a real record manager (not an in mem one)
  • Do less naive retrieval (hybrid semantc + keyword search, include parent information in each chunk, etc)
• Improve prompts, add few-shot examples
• Make sure dbt api calls are correct

LangGraph docs

https://langchain-ai.github.io/langgraph/